add multi-gpu support

[mike-ferguson]

Add multi-GPU and memory-efficient model loading to HuggingfaceSubject

Changes

• Multi-GPU support: When multiple CUDA GPUs are detected, HuggingfaceSubject now loads models with device_map='auto', automatically distributing layers across all available GPUs. Single-GPU and MPS (Apple Silicon) behavior is unchanged.
• Memory-efficient loading: Added low_cpu_mem_usage=True to from_pretrained, reducing peak CPU RAM during checkpoint loading from ~3x to ~1x model size.

Why

Models over ~6B parameters in float32 exceed a single 24 GB GPU. Previously, all weights were loaded onto GPU 0 regardless of how many GPUs were available, causing OOM kills on multi-GPU instances (e.g. g5.12xlarge). This fix allows 7-13B models to run in fp32 on multi-GPU instances without any changes to individual model plugins.

Impact

• No changes required to existing model plugins
• No behavior change on single-GPU or CPU/MPS setups
• Fixes OOM for medium-tier models (Mistral-7B, OPT-6.7B, Falcon-7B, Pythia-12B, etc.) on multi-GPU Batch instances